Feature
MCP server for deployment
Phelix ships a local MCP server that exposes a bounded slice of its machine interface to coding agents. An agent can inspect state and create an immutable plan, but nothing deploys until the plan passes authorization and, when required, a human approval bound to that plan's hash.
simulated
plan created · waiting at the boundary
A transport, not a new engine
The MCP adapter is a local stdio JSON-RPC server. It adds no execution path of its own: every tool call runs the same validation, plan, authorization, approval, and idempotency pipeline as the equivalent CLI command, and returns the same versioned JSON envelope. There is no network listener and no embedded model.
That means the deploy an agent triggers is exactly the deploy you could run by hand, with the same health gate and the same rollback.
simulated
policy
absent · apply denied
no policy on this host · apply is denied by default
Two steps, never one
phelix_plan_create persists a content-addressed plan describing the mutation; it never executes. phelix_plan_apply runs a plan, and only after it passes the full pipeline. There is no single-step deploy, rollback, shell, or filesystem tool.
Approvals bind both the plan id and its content hash, so a human signs off that exact plan, and an agent cannot approve its own action: approving is a separate CLI command, never an MCP tool.
Denied by default
On a host with no authorization policy, reads and plan creation work but applying a plan over MCP is denied. To let an agent execute, the operator writes an enforced policy that allows the mcp caller, ideally requiring approval for each plan.
Once a policy exists, no failure mode degrades into allow-everything: an unreadable or invalid policy fails closed and executes nothing.
- Works with Claude Code, Codex, and any MCP-compatible client over local stdio
- Plan-bound approvals with content hashes
- No network port, no shell tool, no direct deploy tool
- Every refusal is a zero-mutation outcome with a stable error code
Frequently asked questions
Short answers; the docs go deeper.
Can an AI agent deploy to production with Phelix without approval?
No. Agent apply is denied by default, and the recommended enforced policy requires a human approval bound to the exact plan id and content hash before execution.
Which MCP clients work with Phelix?
Any MCP-compatible client over local stdio. Phelix is verified with Claude Code and Codex; you point the client at the phelix binary running mcp serve --stdio.
What can an agent actually do over MCP?
Read context and state, create an immutable plan, and request its apply. Rebuild and rollback happen only through an approved plan on the same pipeline the CLI uses.
One binary between git push and production.
Install the CLI, verify the host, and follow the quick start to build your first application.